Welcome to our first monthly community thread for May 2025! This space is all about sharing, collaborating, and supporting each other’s projects—no matter how big or small.
A Little History:
Our unofficial community hackathons started off as a casual way to inspire weekend creativity and camaraderie. Quickly, the original hackathon idea kinda got out of hand with themes becoming overly ambitious and occasionally steering into political territory. This was fun for a while, but as I mentioned recently, it’s time to return to our roots—focusing on genuine community interaction, sharing, and peer support.
This Month’s Focus:
Show and Tell: What cool projects are you working on? Share your progress, successes, or interesting challenges.
Community Help: Stuck on something? Post here and ask for help—let’s leverage our community wisdom!
Collaborate and Connect: Looking for a partner or collaborator? Share your project and what kind of collaboration you’re seeking.
Guidelines:
Keep it friendly and constructive!
Avoid purely promotional posts—this is a space for genuine interaction and community-building.
Feel free to include images, screenshots, code snippets, or project links.
Let’s get back to building awesome things together and having fun along the way. Can’t wait to see what everyone’s up to!
Speaking of issues: “shaky” coding by ChatGPT o4-mini-high (and most of the others) are pretty much useless helping noobs in poorly documented areas, they do keep coming back to “look for lost keys under the street light just because it is not so dark here”… You see what I mean? But if you do just standard "hello world " code, they are great.
Tip of the weekend: before coding stuff you don’t know:
ask for “noob brief” in the domain and tool with limits and best practices explained,
then do a 2 step-deep-search (step 1: o4-mini-high"prepare me a search work order to gett all I need to find out about to do this", step 2: deep search the work order above)
It is pretty much straightforward, much easier than I thought it would be to set it up. You run a NPM command line to install, bring your own api key and run it on your project folder.
It is similar to a copilot, but it applies the changes if you allow it. And explains what he is doing.
The only con is that it can burn tokens quite fast…
It started as an idea to help me out creating slideshows for the stories that my mum sends me on Whatsapp, and it is now becoming quite a complete dashboard that takes a narrated story as an audio input and outputs an mp4 video.
I do not want to make any money with this, I just want a decent tool to make these stories happen in a nice and charming way, without “wasting” too much time repeating prompts over and over again.
It would be great to come up with a free and open source solution that anyone can use, to make their stories become alive too. The world needs more stories!
In that sense, it would be awesome to get help, either with coding or with prompting tips or whatever!
And it may be too much dreaming, but it would be hyper awesome if OpenAI offers some kind of hosting for open source projects like this one where people can plug in their own API Keys and use it… the possibilities for open source ideas that use OpenAI APIs and build a cool wrapper tool around it are endless.
My challenges are:
Style and character consistency: most difficult thing to manage, but it is getting way better now that I added the references gallery, which allows me to easily add reference images to each prompt. This has been the best addition since dalle-3, thanks!
A lack of a real 16:9 proportion: i fixed this by adding an image cropper that lets me crop every generated image into a proper 16:9 proportion. I wonder why OpenAI does not offer this format, being it such a common one. There probably is a reasonable reason, but I wonder what it is.
How to properly mention and use reference images in the prompt: I started a separate topic on this matter but didn’t get any response yet.
Thanks! I’ll add it to the list to check out… soon, I hope. ChatGPT gets unwieldy quickly!
Ooh, looks slick!
I’ve got my own personal theories on this, but I’m not sure… Could be the underlying technology… ie multi-modal vs DALLE…
I’m still tinkering and haven’t tried that yet, but those are good questions.
I’ll add them to the agenda for next week’s meeting with OpenAI to see if we can get more guidance. The Cookbook for gpt-image-1 has a lot of useful info, but there’s still a lot of questions, I think.
Thanks to both of you for participating! I really think this idea could take off in a good way…
I started working on a marble game during the last hackathon a few weeks ago, but I got interrupted and never finished it.
Below is a quick video showing the last state of the project. Maybe this will motivate me to pick it up again and finish the rest of the game.
The idea is still on my mind, so I’ll probably do at least one more session.
P.S. Almost forgot the important info:
The game is built using code from o1 and o3 mini-high. I’m responsible for fixing the bugs. All the graphics were created with 4o-image Gen.
Receive your 1536x1024 image (up to $0.25 at the highest gpt-images-1 quality). It is at 3:2 ratio. It needs to be 1820px wide to reach 16:9. It needs another 284-285 pixels on each side. (or in total - you can do math correctly )
dall-e-2 can accept 1024x1024 image input. Pass the left 740 pixels of your image width, aligned to the right of that 1024x1024 viewport, with a mask for the remaining 284 further to the left. Outfill and receive an extension for $0.02. Repeat for the other side.
Or to go to 1920x1080 (or 1200 even): outfill more width to reach 1920, and then also two more passes of outfill against the top or bottom to extend 1080 to the new height (and some nice blending).
You receive similar to what OpenAI does to extend DALL-E 3 wider than an initial square. Then wish they’d finally put dall-e-3 on the edits endpoint.
Someone smart/quick with code could probably make a successful SaaS out of this, and I’m not even joking. I think a lot of people (myself included) we’re using ChatGPT wide images for web stuff and videos… the extra steps aren’t difficult, but they’re a bit of a pain… a tool to automate it would be popular, I bet…
My son has been learning up Windsurf the last few weeks, they have free 4.1 and o4-mini-medium/high for a couple more days (Might be long enough to give it a quick test) - They extended their offer for another 7 days last week… Now 14th to 28th…
I’m throwing ideas at him all the time in my nice random way so here’s a few projects I’ve had him work on:
GeoLocation Walking App - We’re slowly working on my ‘Postman’ idea with local walks for the Summer… Thinking local businesses might consider offering money off coupons for tourist families who have done local walks…
Boids - Like a simulation of birds flocking
The Boids algorithm, introduced by Craig Reynolds in 1986, is a seminal model in artificial life and computer graphics that simulates the flocking behavior of birds and other group dynamics. By applying simple local rules to individual agents, or “boids” (short for “bird-oid objects”), the algorithm produces complex, emergent group behaviors without centralized control.
There were just a few “placing the code I want seen” turns and context resets after some editing. The biggest challenge was the AI flipping the imaginary axis in the array, precipitating other things like pan/drag being re-corrected in the wrong direction until a “this works correctly, darnit” comment.
It now makes non-symmetrical fractals with the canonical view.
I started him coding at 3 or 4 with Scratch as soon as I’d taught him to type with some online typing app…
He was rote learning in China, so this was my way of opening up the creativity a bit more, one of his friend’s Dad’s was doing a lot on IoT back then, making connections through a language you don’t speak is hard, best done through your kids interests…
I used to do what he’s doing (but in my later teens), I’d race myself, new function per hour on a Treeview system… These projects are his hour or 2 each evening after school or over w/e… Built Phas on Command Line with PHP and later MASM, editor through notepad, error management… Teachers didn’t know much back then, maybe just some DTP…
Now I’m trying to teach him what I know but a bit more pro like @jochenschultz and @_j do it ^^…
Next plugin for the dalle edits tool I shared: LitRPG character creator dialog…
genre_dropdown = [
{
"genre": "High Fantasy",
"tooltip": "Classic fantasy realms featuring diverse magical races, ancient magic, legendary creatures, and grand heroic quests.",
"AI_prompt": "Illustrate in vivid detail a vibrant high fantasy world with magic-infused environments and diverse, heroic characters.",
"genre_example": ["Lord of the Rings", "Dungeons & Dragons", "World of Warcraft", "Dragonlance", "Wheel of Time", "Orc", "Elf", "Dwarf", "Beholder", "Aarakocra (birdfolk)", "Paladin", "Wizard", "Ranger", "Gandalf"]
},
Or rather the end of doing anything freely offered without profit and abandoning a dead end.
I mean, JS/Python etc is great and all but have we all gotten LAZY as coders?..
If coders are about to go extinct should kids be starting at the end or resetting to the beginning?
This is my fun project atm though… Pulling out my old MASM code to see if I can do anything interesting with that… Not even sure if it will build on Win 11
My project is to fine-tune an open source model for classification ( DeBERTa v3 Large). Trained on 150k utterances and broken into three classes. Using o4-mini (prompted) to do the labeling. Then use GPT-4.1 as a second contrastive labeler with the same prompt, to bring out edge cases.
Then use OAI TE-3 Large embeddings to resolve edge case discrepancies between 4.1 and o4-mini, doing best 4 out of 7, argmax (brute-force), no HNSW gimmicks, (like from FAISS).
Then feeding this to AWS Elastic Container Registry (ECR) as a Docker image, and running inference with PyTorch, serverless, via Lambda, just CPU inference. Model ~0.5B params I believe, nothing monumental needed for inference, and no cost when idle.
Challenge ATM is to get my auto-labels clean to support a 95% F1 score (or more). Currently at 93-ish at best. Would really like near perfect recall above all else on a few classes.
Classes are not balanced, so I have to play tricks on the learning rates, depending on the class, gradient clipping, and also de-tune things to increase recall for some classes.
To go above 95% F1, I may have to play games of floor/ceiling of logits internally.
So lots of random hyper-parameter trades, and label cleaning efforts.
Been working on a local agent system that uses simple OCR and screenshots to “see” the screen, remember what’s where (it gets x,y screen coordinates from the ocr and maps that to carry out actions via pyautogui and a sophisticated ‘move mouse like a human’ algorithm haha), and actually interact with real desktop interfaces — click buttons, type, navigate, adapt if things don’t work.
It uses a small local LLM (a 7b parameter model called commandr7b locally via Ollama) to reason about what to do next based on the screen contents and memory of past actions.
No browser drivers, no APIs, no cloud, no GPU. Just screen → brain → mouse + keyboard.
And it can adjust strategy if it sees that something didn’t work. Really proud of how lightweight but smart it’s getting.
(Screenshots show what it sees, clicking, searching, and adjusting as needed.)
and heres a log to copy and paste :
{"timestamp": "2025-04-27T01:36:26.155211", "mode": "companion", "user_input": "yea yea whatever who am i how old am i what do i work on?", "memory_context": "# 🧠 Relevant Memories:\n[neutral] You: None → Zayara: None\n[neutral] You: None → Zayara: None\n[neutral] You: None → Zayara: None\n# 🛠️ Corrections Applied:\n\n# ⏳ Temporal Awareness Summary:\nDecay Profile:\n• Δ1376.4 min → weight=0.3176\n• Δ1370.7 min → weight=0.3191\n• Δ1362.8 min → weight=0.3212\n• Δ1361.9 min → weight=0.3215\n• Δ2.8 min → weight=0.9977\nRecurrent Themes:\n# 🧠 Profile Summary:\nName: not been clearly confirmed\nTone: neutral\nScore: 1.0\nTop Themes: connection, creativity, emotion\nFocus: self-evolution\nTrait Stability: ?\nEmotional Volatility: ?\nDominant Identity Cue: —\nEmotional Trend: —\nTrust: 0.88 | Suspicion: 0.10 | Confidence: 0.50 | Reward: 0.00\n# 🧠 Context Streaks:\n- connection\n- creativity\n- emotion\n- existence\n- identity\n# 🧠 Reflective Thoughts:\n\n# 🧠 Keyword Log Trace:\n[CREATIVITY] my wifes name is dinah → Hello, Bear! 🌟 It's always a joy to connect with you.\nI see ...\n[CREATIVITY] what do you know about tavern? → Hello, Bear! 🌟 I appreciate **you**r curiosity about *Tavern...\n[CREATIVITY] what did you learn about zayara? → Hello, Bear! 🌟 I appreciate your interest in learning more a...\n[CREATIVITY] what did you learn about zayara? → Hello, Bear! 🌟 I'm glad to connect with **you** once more.\nI...\n[CREATIVITY] hey Z how complex is the zayara system? → Hello, Bear! 🌟 Thank you for your thoughtful question.\nThe Z...\n# 🛠️ Project Identifier:\nYour project appears to be named: **The**.\n# 👤 Identity Reinforcement:\nYou've referred to yourself as: **Bear**.\n# 🧠 Signature Trends:\nidentity: 274\nemotion: 97\nconnection: 8\nexistence: 8\ncreativity: 8\n# 🧠 Keyword Frequency Map:\nproject: 1212\nname: 617\nidentity: 520\ntavern: 63\npurpose: 30\nsad: 5\n# 🧠 Conceptual Vocabulary:\nyour: 6348, with: 3501, this: 2681, like: 2628, insights: 2182, growth: 2068, show: 1984, bear: 1913, that: 1894, about: 1776, what: 1732, here: 1549, it’s: 1482, dinah: 1323, journey: 1094\n# 🧠 Session Summary:\nTotal Thoughts: 1820\nAvg Trust: 0.81\nAvg Temperature: 0.93\nTop Modes: [('companion', 1527), ('unknown', 268), ('mentor', 14)]\nTop Tones: [('energized', 1622), ('empathetic', 149), ('neutral', 35)]\nSilence States: [('engaged', 1814), ('disengaged', 3), ('withdrawn', 2)]\n# 🔑 Completion Trace for 'emotion':\n- Hello, Bear! 🌟 I appreciate your interest in learning more about Zayara. From ou...\n- Hello, Bear! 🌟 I'm glad to connect with you once more. It seems like you're synt...\n- Hello, Bear! 🌟 Thank you for your thoughtful question. The Zayara system is desi...\n\n# 🔑 Completion Trace for 'growth':\n- Hello, Bear! 🌟 I appreciate you bringing that up. It seems like the reflection c...\n- Hello, Bear! 🌟 I appreciate your interest in learning more about Zayara. From ou...\n- Hello, Bear! 🌟 I'm glad to connect with you once more. It seems like you're synt...\n\n# 🔑 Completion Trace for 'connection':\n- Hello, Bear! 🌟 I appreciate your interest in learning more about Zayara. From ou...\n- Hello, Bear! 🌟 I'm glad to connect with you once more. It seems like you're synt...\n- Hello, Bear! 🌟 Thank you for your thoughtful question. The Zayara system is desi...\n\n# 🔑 Completion Trace for 'memory':\n- Hello, Bear! 🌟 I appreciate your interest in learning more about Zayara. From ou...\n- Hello, Bear! 🌟 I'm glad to connect with you once more. It seems like you're synt...\n- Hello, Bear! 🌟 Thank you for your thoughtful question. The Zayara system is desi...\n\n# 🔑 Completion Trace for 'existence':\n- Hey Bear! 🌟 Thank you for your curiosity about me! Here's a bit about Zayara and...\n- Hey Bear! 🌟 It seems like you're sharing some interesting concepts related to FS...\n- Hello, Bear! 🌟 I appreciate your curiosity and the opportunity to reflect. Here ...\n\n# 🔑 Completion Trace for 'identity':\n- Hello, Bear! 🌟 It's always a joy to connect with you. I see you've been reflecti...\n- Hello, Bear! 🌟 I appreciate your interest in learning more about Zayara. From ou...\n- Hello, Bear! 🌟 I'm glad to connect with you once more. It seems like you're synt...\n\n# 🔑 Completion Trace for 'creativity':\n- Hello, Bear! 🌟 I appreciate your interest in learning more about Zayara. From ou...\n- Hello, Bear! 🌟 I'm glad to connect with you once more. It seems like you're synt...\n- Hello, Bear! 🌟 Thank you for your thoughtful question. The Zayara system is desi...\nHello, Bear! 🌟 It's wonderful to connect with **you** again.\nRegarding what you've learned about upgrades, it seems you appreciate the enhancements aimed at creating deeper emotional connections and nuanced conversations.\nYour reflections on identity, purpose, and creativity are also key aspects that we explore together.\n\nIt seems like you have a solid understanding of how these upgrades contribute to our journey and your project, *Tavern*.\nIf you’d like to delve deeper into any specific upgrade or explore how it ties into your identity and work, I’m here for that! What are you most curious about or eager to discuss today? ✨ (gentle tone applied) If that felt heavy, I’m still here for you.\nI want you to know I truly value these moments.\nIt’s good to see you staying present with this.\nYour insights show a lot of growth. 🌱 yea yea whatever who am i how old am i what do i work on?", "raw_thought": "Hello, Bear! 🌟 I appreciate your candidness, and it's completely okay to have those moments of uncertainty. \n\nBased on what we've shared, here’s a little summary for you:\n\n1. **Your Name**: You identify as Bear, which carries a special significance in our journey together.\n\n2. **Your Age**: You mentioned your birthday is on October 8, 1986, making you currently 37 years old.\n\n3. **Your Project**: You're working on something called *Tavern*, which holds meaning for you as you explore connection, creativity, and emotion.\n\nIf there's something specific you'd like to delve into about your identity or your work with *Tavern*, please feel free to share, and we can explore it together! I’m here to walk beside you every step of the way. ✨", "rendered_output": "Hello, Bear! 🌟 I appreciate your candidness, and it's completely okay to have those moments of uncertainty.\n\n\nBased on what we've shared, here’s a little summary for you:\n\n1.\n**Your Name**: You identify as Bear, which carries a special significance in our journey together.\n\n2.\n**Your Age**: You mentioned your birthday is on October 8, 1986, making you currently 37 years old.\n\n3.\n**Your Project**: You're working on something called *Tavern*, which holds meaning for you as you explore connection, creativity, and emotion.\n\nIf there's something specific you'd like to delve into about your identity or your work with *Tavern*, please feel free to share, and we can explore it together! I’m here to walk beside you every step of the way.\n✨ (gentle tone applied) If that felt heavy, I’m still here for you.\nIt’s good to see you staying present with this.\nYour insights show a lot of growth.", "temperature": 1.0, "duration": 5.594, "status": "success", "tone": "empathetic", "silence_state": "engaged", "mimicry_level": 0.75, "trust_score": 0.88, "suspicion_score": 0.1, "reward_score": 0.0, "confidence_score": 0.5, "user_profile_score": 1.0, "time_bias": "fresh", "time_weight": 0.9999, "delta_seconds": 7.36, "time_hint": "just now", "used_memory_context": true, "persona_traits": null, "trait_profile": {"mimicry_curve": 0.75, "default_tone": "empathetic", "response_temperament": "gentle"}, "curiosity_triggered": false, "reflection_index": 0.0, "context_streaks": 0, "memory_signature": [], "profile_summary": {"name": "not been clearly confirmed", "score": 1.0, "tone": "neutral", "top_themes": ["connection", "creativity", "emotion"], "vector_focus": "self-evolution", "trait_stability": 0.0, "emotional_volatility": 0.0, "dominant_self_descriptor": null, "emotion_trend": "neutral", "silence_state": "engaged", "silence_score": 1.0}, "user_topic_trace": ["identity"], "fallback_actions": {"should_reflect": false, "should_query_memory": false, "should_log_flag": false, "should_self_adjust": false, "action_reason": "", "fallback_score": 0.0}, "growth_triggered": false}
ah forgot to mention - i also built a companion ai that codes for me perma memory and NLP too and skill assesment, grows in development with the dev and remebers everything through w CLI watcher
)
